Question

I want to determine the column classes of a large data.table.

colClasses <- sapply(DT, FUN=function(x)class(x)[1])

works, but apparently local copies are stored into memory:

> memory.size()
[1] 687.59
> colClasses <- sapply(DT, class)
> memory.size()
[1] 1346.21

A loop seems not possible, because a data.table “with=FALSE” always results in a data.table.

A quick and very dirty method is:

DT1 <- DT[1, ]
colClasses <- sapply(DT1, FUN=function(x)class(x)[1])

What is the most elegent and efficient way to do this?

Answer 1

Have briefly investigated, and it looks like a data.table bug.

> DT = data.table(a=1:1e6,b=1:1e6,c=1:1e6,d=1:1e6)
> Rprofmem()
> sapply(DT,class)
        a         b         c         d 
"integer" "integer" "integer" "integer" 
> Rprofmem(NULL)
> noquote(readLines("Rprofmem.out"))
[1] 4000040 :"as.list.data.table" "as.list" "lapply" "sapply"       
[2] 4000040 :"as.list.data.table" "as.list" "lapply" "sapply" 
[3] 4000040 :"as.list.data.table" "as.list" "lapply" "sapply"   
[4] 4000040 :"as.list.data.table" "as.list" "lapply" "sapply" 

> tracemem(DT)
> sapply(DT,class)
tracemem[000000000431A290 -> 00000000065D70D8]: as.list.data.table as.list lapply sapply 
        a         b         c         d 
"integer" "integer" "integer" "integer" 

So, looking at as.list.data.table :

> data.table:::as.list.data.table
function (x, ...) 
{
    ans <- unclass(x)
    setattr(ans, "row.names", NULL)
    setattr(ans, "sorted", NULL)
    setattr(ans, ".internal.selfref", NULL)
    ans
}
<environment: namespace:data.table>
> 

Note the pesky unclass on the first line. ?unclass confirms that it takes a deep copy of its argument. From this quick look it doesn’t seem like sapply or lapply are doing the copying (I didn’t think they did since R is good at copy-on-write, and those aren’t writing), but rather the as.list in lapply (which dispatches to as.list.data.table).

So, if we avoid the unclass, it should speed up. Let’s try:

> DT = data.table(a=1:1e7,b=1:1e7,c=1:1e7,d=1:1e7)
> system.time(sapply(DT,class))
   user  system elapsed 
   0.28    0.06    0.35 
> system.time(sapply(DT,class))  # repeat timing a few times and take minimum
   user  system elapsed 
   0.17    0.00    0.17 
> system.time(sapply(DT,class))
   user  system elapsed 
   0.13    0.04    0.18 
> system.time(sapply(DT,class))
   user  system elapsed 
   0.14    0.03    0.17 
> assignInNamespace("as.list.data.table",function(x)x,"data.table")
> data.table:::as.list.data.table
function(x)x
> system.time(sapply(DT,class))
   user  system elapsed 
      0       0       0 
> system.time(sapply(DT,class))
   user  system elapsed 
   0.01    0.00    0.02 
> system.time(sapply(DT,class))
   user  system elapsed 
      0       0       0 
> sapply(DT,class)
        a         b         c         d 
"integer" "integer" "integer" "integer" 
> 

So, yes, infinitely better.

I’ve raised bug report #2000 to remove the as.list.data.table method, since a data.table is() already a list, too. This might speed up quite a few idioms actually, such as lapply(.SD,...). [EDIT: This was fixed in v1.8.1].

Thanks for asking this question!!